Model Fitting and Evaluation

Import Dataset

# import dataset
data_clean <- read_csv("data/data-clean.csv")

# quick check
head(data_clean, 10)

Cross-Validation Scheme

# set seed
set.seed(100)

# create initial split
splitted <- initial_split(data_clean, prop = 0.8, strata = "sentiment")

# quick check
splitted

#> <11681/2919/14600>

Defining Preprocess Recipe

# define preprocess recipe from train dataset
rec <- recipe(sentiment ~ ., data = training(splitted)) %>% 
  step_rm(-sentiment, -tweet) %>%
  step_string2factor(sentiment, levels = c("negative", "neutral", "positive")) %>%
  step_downsample(sentiment, ratio = 1/1, seed = 100) %>%
  step_mutate(tweet = str_squish(tweet)) %>% 
  step_mutate(tweet = replace_html(tweet, symbol = FALSE)) %>% 
  step_mutate(tweet = replace_kern(tweet)) %>% 
  step_mutate(tweet = replace_word_elongation(tweet)) %>% 
  step_mutate(tweet = replace_date(tweet, replacement = "datewords")) %>% 
  step_mutate(tweet = replace_time(tweet, replacement = "timewords")) %>% 
  step_mutate(tweet = replace_money(tweet, replacement = "moneywords")) %>% 
  step_mutate(tweet = replace_ordinal(tweet, remove = FALSE)) %>% 
  step_mutate(tweet = replace_number(tweet, remove = FALSE)) %>% 
  step_mutate(tweet = replace_internet_slang(tweet)) %>% 
  step_mutate(tweet = replace_contraction(tweet)) %>% 
  step_mutate(tweet = replace_emoji(tweet)) %>% 
  step_mutate(tweet = replace_symbol(tweet)) %>% 
  step_mutate(tweet = str_squish(tweet)) %>% 
  step_mutate(tweet = str_replace_all(tweet, "(<.*>)", "")) %>% 
  step_mutate(tweet = str_replace_all(tweet, "[:digit:]", "")) %>% 
  step_tokenize(tweet, token = "words") %>%
  step_stem(tweet) %>%
  step_stopwords(tweet) %>%
  step_tokenfilter(tweet, max_tokens = 256) %>%
  step_tfidf(tweet) %>%
  prep(string_as_factor = FALSE)

# get train and test dataset
data_train <- juice(rec)
data_test <- bake(rec, testing(splitted))

# quick check
head(juice(rec), 10)

Defining Model Specifications

# define model specification
model_spec <- rand_forest(
  mode = "classification",
  mtry = 3,
  trees = 500,
  min_n = 1
)

# define model engine
model_spec <- set_engine(
  object = model_spec,
  engine = "ranger",
  seed = 100,
  num.threads = parallel::detectCores() / 2,
  importance = "impurity"
)

# quick check
model_spec

#> Random Forest Model Specification (classification)
#> 
#> Main Arguments:
#>   mtry = 3
#>   trees = 500
#>   min_n = 1
#> 
#> Engine-Specific Arguments:
#>   seed = 100
#>   num.threads = parallel::detectCores()/2
#>   importance = impurity
#> 
#> Computational engine: ranger

Model Fitting

# fit the model
model <- fit_xy(
  object = model_spec,
  x = select(data_train, -sentiment),
  y = select(data_train, sentiment)
)

# quick check
model

#> parsnip model object
#> 
#> Ranger result
#> 
#> Call:
#>  ranger::ranger(formula = formula, data = data, mtry = ~3, num.trees = ~500,      min.node.size = ~1, seed = ~100, num.threads = ~parallel::detectCores()/2,      importance = ~"impurity", verbose = FALSE, probability = TRUE) 
#> 
#> Type:                             Probability estimation 
#> Number of trees:                  500 
#> Sample size:                      5661 
#> Number of independent variables:  256 
#> Mtry:                             3 
#> Target node size:                 1 
#> Variable importance mode:         impurity 
#> Splitrule:                        gini 
#> OOB prediction error (Brier s.):  0.2962273

Variable Importance

# get variable importance
var_imp <- tidy(model$fit$variable.importance)

# tidying
var_imp <- var_imp %>%
  arrange(desc(x)) %>% 
  head(10) %>% 
  rename(variable = names, importance = x) %>%
  mutate(variable = reorder(variable, importance))

# variable importance plot
ggplot(var_imp, aes(x = variable, y = importance)) +
  geom_col(fill = "darkblue") +
  coord_flip() +
  labs(title = "Variables Importance (Top 10)", x = NULL, y = NULL, fill = NULL) +
  scale_y_continuous(expand = expand_scale(mult = c(0, 0.1))) +
  theme_minimal()

Predict on Test Dataset

# predict on test
pred_test <- select(data_test, sentiment) %>%
  bind_cols(predict(model, select(data_test, -sentiment))) %>%
  bind_cols(predict(model, select(data_test, -sentiment), type = "prob"))

# quick check
head(pred_test, 10)

Confusion Matrix

# confusion matrix
pred_test %>%
  conf_mat(sentiment, .pred_class) %>%
  autoplot(type = "heatmap")

# metrics summary
pred_test %>%
  summarise(
    accuracy = accuracy_vec(sentiment, .pred_class),
    sensitivity = sens_vec(sentiment, .pred_class),
    specificity = spec_vec(sentiment, .pred_class),
    precision = precision_vec(sentiment, .pred_class)
  )

ROC Curve

# plot roc curve
pred_test %>%
  roc_curve(sentiment, .pred_negative:.pred_positive) %>%
  autoplot()

# get roc curve data on test dataset
pred_test_roc <- pred_test %>%
  roc_curve(sentiment, .pred_negative:.pred_positive)

# tidying
pred_test_roc <- pred_test_roc %>% 
  mutate_if(~ is.numeric(.), ~ round(., 4)) %>% 
  gather(metric, value, -.threshold, -.level)

# plot sensitivity-specificity trade-off
p <- ggplot(pred_test_roc, aes(x = .threshold, y = value)) +
  geom_line(aes(colour = metric)) +
  facet_wrap(~ .level, ncol = 1) +
  labs(x = "Probability Threshold to be Classified as Positive", y = "Value", colour = "Metrics") +
  theme_minimal()

ggplotly(p)

Precision-Recall Curve

# plot pr curve
pred_test %>%
  pr_curve(sentiment, .pred_negative:.pred_positive) %>%
  autoplot()

# get pr curve data on test dataset
pred_test_pr <- pred_test %>%
  pr_curve(sentiment, .pred_negative:.pred_positive)

# tidying
pred_test_pr <- pred_test_pr %>% 
  mutate_if(~ is.numeric(.), ~ round(., 4)) %>% 
  gather(metric, value, -.threshold, -.level)

# plot recall-precision trade-off
p <- ggplot(pred_test_pr, aes(x = .threshold, y = value)) +
  geom_line(aes(colour = metric)) +
  facet_wrap(~ .level, ncol = 1) +
  labs(x = "Probability Threshold to be Classified as Positive", y = "Value", colour = "Metrics") +
  theme_minimal()

ggplotly(p)